Title card reading 'WeMM-Video-9B vs Hailuo 2.3' with the subtitle 'Six silent seconds', two panels split by a vs disc: the left headed WeMM-Video-9B with an 'open weights' badge and rows reading '19.54 GiB BF16', 'reads 16 frames at 2 fps', '8 seconds of coverage', 'accepts audio input', 'no per-clip cost'; the right headed Hailuo 2.3 with a 'metered API' badge and rows reading '6 s at 1080p', '10 s at 768p', 'silent output', 'no end-frame control', 'about $0.48 per clip'.
Guides & Insights

WeMM-Video-9B vs Hailuo 2.3: Six Silent Seconds

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Both of these models will let you down in exactly one place, and it is the same place. Hailuo 2.3 produces six seconds of 1080p video and no sound at all. WeMM-Video-9B reads sixteen sampled frames of video, which at the shipped sampling rate is about eight seconds of coverage, and returns text. Neither one is a general-purpose video system, and neither is on a leaderboard you can point at. M​iniMax shipped Hailuo 2.3 in April 2026 and has since moved on to a model that does appear on those boards; T​encent's WeChat Vision group put WeMM-Video-9B on Hugging Face on 17 November 2026 with no announcement, no endpoint and no price. That leaves two releases that are each, in their own way, a bounded tool — and the boundary is the interesting part.

A note on sourcing before anything else, because this piece leans on it. The Hailuo 2.3 numbers come from M​iniMax's own pricing and documentation plus third-party per-second trackers, since the model has no independent Elo. The WeMM-Video-9B numbers come from the repository manifests, the model card and arXiv 2611.26409. Nothing here has been reproduced outside the labs that published it, and where a figure is vendor-reported the text says so.

The six seconds are not a default

Hailuo 2.3's duration limit is a trained property, not a setting. Six seconds at 1080p. Ten seconds if you drop to 768p. There is no end-frame conditioning, so you cannot ask it to land on a specific final composition the way you can with a keyframe model, and there is no native audio track at all — the output is silent, and any sound you want has to be added downstream or generated separately.

Pricing runs through prepaid points rather than a published per-second rate, which makes cost comparisons harder than they should be. Six seconds at 768p on the standard tier is one point; ten seconds at 768p or six seconds at 1080p is two. The Fast variant lands around 0.7 points for the short case and 1.1 to 1.3 for the long one. Third-party trackers that have converted this to dollars put it near $0.08 per second, or roughly $0.48 for a six-second clip.

• Duration — Hailuo 2.3: 6 s at 1080p, 10 s at 768p. WeMM-Video-9B: no output duration, because it produces no video.

• Audio — Hailuo 2.3: none, output is silent. WeMM-Video-9B: a 1.08B Whisper-style tower that transcribes and reasons over the soundtrack you supply.

• Control — Hailuo 2.3: no end-frame conditioning. WeMM-Video-9B: 4 to 128 frames configurable, 2 fps default sampling.

Where Hailuo 2.3 is genuinely good is the thing its users talk about: cinematic camera movement, human motion and micro-expression, and a stylised, animation-adjacent look that holds up better than most models on non-photoreal briefs. That is a real strength and it is not measurable on a board, which is part of why the model has never appeared on one.

Two-column scoreboard headed 'WeMM-Video-9B vs Hailuo 2.3 — the scoreboard', six shared dimensions in rows: Price ('no per-clip cost' against 'about $0.48 per clip'), Board position ('none, unrankable' against 'not on AA boards'), Duration ('8 s of sampled input' against '6 s at 1080p, 10 s at 768p'), Audio ('reads the soundtrack' against 'silent output'), Control ('16 frames at 2 fps' against 'no end-frame control'), Release ('17 Nov 2026' against 'April 2026'), with a footer reading 'Hailuo pricing per October 2026 vendor rates; WeMM-Video-9B figures vendor-reported.'

The board has no row for either

Read on 20 November 2026, all three Artificial Analysis video boards — AA-Video-T2V v2.1, AA-Video-I2V v2.1, AA-Video-Editing v1.1 — return nothing for Hailuo 2.3. Not a low rank. No entry.

That absence has two different causes and it is worth separating them. WeMM-Video-9B is unrankable: the boards score generated footage by pairwise preference, and a model that outputs tokens has nothing to submit. Hailuo 2.3 could in principle be ranked and has not been — 1080p output that a rater could watch, no submission, no price row, no sample count.

So the only quality evidence in this matchup sits on the side with no product. The WeMM-Video-9B paper reports, in thinking mode on joint audio-visual benchmarks, wins over Qwen3-Omni-30B-A3B-Instruct on 35 of 37 shared figures — Video-MME with audio 85.1 against 77.3, Daily-Omni 83.6 against 70.7, EvaOmni 50.7 against 26.1 — and one loss, FutureOmni at 53.7 against 55.1. Against its own Qwen3.5-9B backbone on silent video it gains on F-16-NBA and TimeLens and loses on MLVU, LVBench, VideoVista and MVBench. All of that is the authors' own harness, disclosed as such in the paper.

Screenshot of the arXiv abstract page for 2611.26409, 'Pushing the Frontiers of Open-Source Omni-Modal Video Understanding', marked submitted 19 November 2026, with the author list and the abstract describing the WeMM-Video family as 2B, 4B, 9B and 35B-A3B variants.

The generation after Hailuo 2.3

Hailuo 2.3 is not M​iniMax's current model, and pretending otherwise would be a disservice to anyone budgeting this quarter. The successor, M​iniMax H3, was released on 31 July 2026 and does appear on the boards.

• AA-Video-T2V v2.1 — M​iniMax H3 at 768p ranks 3rd on Elo 1,161 ±13 over 8,386 samples, listed July 2026 at $4.80 per minute.

• AA-Video-I2V v2.1 — 5th at Elo 1,134 ±14 over 6,594 samples, same rate.

• AA-Video-Editing v1.1 — 6th at Elo 1,126 ±10 over 9,688 samples, $9.60 per minute.

Third on text-to-video and fifth on image-to-video at $4.80 a minute is a different conversation from Hailuo 2.3's unranked, point-priced, six-second offering. If you are choosing a M​iniMax model rather than comparing one, H3 has already won that argument inside the vendor's own lineup.

Screenshot of the Artificial Analysis text-to-video leaderboard, AA-Video-T2V v2.1, with ranked rows carrying Elo, confidence interval, sample counts, release month and per-minute price, headed by Wan 3.0 at 1,176 and showing MiniMax H3 at 768p third on 1,161 over 8,386 samples at $4.80 per minute; no Hailuo 2.3 row appears.

Why the eight seconds matter more than they look

The shipped WeMM-Video-9B recipe overrides its own generous processor defaults — 2 fps, minimum 4 frames, maximum 128, a 100,663,296-pixel budget — down to 2 fps and 16 frames at 262,144 pixels, sampled around the midpoint of the clip. Sixteen frames at two per second is eight seconds of coverage, and the coverage starts in the middle.

Put that next to Hailuo 2.3's six seconds and something clarifying happens. Hailuo 2.3 produces less footage than WeMM-Video-9B can look at, in one sense: a six-second Hailuo clip fits inside the describer's window with room to spare. But the ten-second 768p mode does not, and neither does anything you would generate with a different model and then hand to WeMM-Video-9B for review. The video side of the pipeline caps out around eight seconds while the audio side can hold roughly thirteen minutes at ten tokens per second, and every serving command in the repository passes an 8,192-token context length that makes that asymmetry the binding constraint.

• If you are generating — Hailuo 2.3 gives you six seconds, silent, with strong camera motion and human detail.

• If you are reviewing — WeMM-Video-9B gives you eight seconds of sampled frames plus the full audio track, described in English or Chinese.

• The join — a six-second silent generation is inside both windows, which is the one configuration where this pairing is clean.

Failover is the part that is not theoretical

Neither model is served through OrcaRouter, and that is stated here rather than buried. Hailuo 2.3 is reachable through M​iniMax's own platform and several third-party platforms; WeMM-Video-9B is reachable by downloading 19.54 GiB of Apache-2.0 weights.

What matters for a pipeline that mixes a generator and a describer is what happens when one leg drops. OrcaRouter's failover routes around a model that is rate-limited or returning errors without the caller rewriting its client, and the routing DSL lets you express that as a policy rather than a pile of retry logic. For a workload where the describing step is a local checkpoint and the generating step is a hosted API, the hosted half is the half that goes down — and having it fail over to a comparable model instead of failing the job is worth more than a marginal price difference.

Six seconds is either enough or it is not

If six seconds of silent, cinematically confident 1080p is the deliverable, Hailuo 2.3 does that and the only real question is whether the point pricing beats the per-second alternatives for the volume you need. Check the ten-second 768p tier before you standardise on the short one; the point cost doubles, not the per-second rate.

If the deliverable is a written account of footage, WeMM-Video-9B does that with the strongest audio-visual numbers in the family and no independent confirmation of any of them. Budget for a patched vLLM or SGLang install and for the fact that you will be the first person outside T​encent to run it in anger.

What neither one is, is a general video platform. They are two narrow tools with an eight-second and a six-second edge, and picking between them is a question about which edge you can afford to hit.